Models

Models with optimized parameters

Embedding	Epochs	Hidden size	Learning rate	Mini batch size	rnn layers	f1-score	precision	recall	accuracy
GloVe	150	32	0.15	16	2	0.7685	0.7706	0.7665	0.6241
fastText	150	256	0.1	16	2	0.8042	0.8351	0.7755	0.6725
GloVe and fastText	150	256	0.1	16	2	0.8334	0.8632	0.8056	0.7144
GloVe and fastText	150	256	0.1	16	4	0.8319	0.8616	0.8041	0.7121
BERT

Models with different learning rates

Fixed Parameters:

Epochs	Hidden size	Mini batch size	rnn layers
150	256	16	2

Models

Embedding	Learning r.	f1-score	precision	recall	accuracy	zor_cumleler accuracy
GloVe(50)	0.05	0.8031	0.8217	0.7853	0.6709
GloVe(50)	0.1	0.7973	0.8379	0.7605	0.6630
GloVe(50)	0.15	0.7885	0.8276	0.7530	0.6509
GloVe(50)	0.2	0.7866	0.8306	0.7470	0.6482
GloVe(300)	0.2	0.8422	0.8709	0.8153	0.7274	0.55
fastText	0.05	0.8178	0.8361	0.8003	0.6918
fastText	0.1	0.8042	0.8351	0.7755	0.6725
fastText	0.15	0.8115	0.8448	0.7808	0.6829
fastText	0.2	0.8323	0.8705	0.7973	0.7128	0.64
word2vec	0.2	0.8339	0.8767	0.7950	0.7151	0.71
word2vec and fastText	0.2	0.8364	0.8759	0.8003	0.7188	0.69
GloVe(50) and fastText	0.05	0.8266	0.8471	0.8071	0.7045
GloVe(50) and fastText	0.1	0.8334	0.8632	0.8056	0.7144
GloVe(50) and fastText	0.15	0.8355	0.8625	0.8101	0.7174
GloVe(50) and fastText	0.2	0.8409	0.8769	0.8078	0.7256
GloVe(300) and fastText	0.05	0.8264	0.8542	0.8003	0.7041
GloVe(300) and fastText	0.1	0.8374	0.8648	0.8116	0.7202
GloVe(300) and fastText	0.15	0.8595	0.9116	0.8131	0.7537
GloVe(300) and fastText	0.2	0.8605	0.9055	0.8198	0.7552	0.66
GloVe(300) and word2vec	0.2	0.8563	0.8979	0.8183	0.7486	0.67
GloVe(300) and fastText and word2vec	0.1	0.8548	0.8983	0.8153	0.7464
GloVe(300) and fastText and word2vec	0.15	0.8597	0.9074	0.8168	0.7540
GloVe(300) and fastText and word2vec	0.2	0.8667	0.9156	0.8228	0.7648	0.65

Zor Cumleler Accuracy

Embedding	Learning r.	zor_cumleler accuracy
GloVe(300)	0.2	0.55
fastText	0.2	0.64
word2vec	0.2	0.71
GloVe(300) and fastText	0.2	0.66
GloVe(300) and word2vec	0.2	0.67
fastText and word2vec	0.2	0.69
GloVe(300) fastText and word2vec	0.2	0.65
Google Docs		0.34
Microsoft Office		0.29
LibreOffice		0.0
ITU NLP Pipeline		0.0

Comparison Methods

Those are the methods that don't use semantic analysis but follows the same syntactic rule.

Model 1: Always disjoint 'de/da', if the word ends with 'de' or 'da' mark it as having error

Model 2: Always joint 'de/da', if the word is 'de' or 'da' mark it as having error

Model	True positive	True negative	False positive	False negative	f1-score	precision	recall
Model 1	3521	6828	29692	10194	0.15	0.1060	0.2567
Model 2	10194	29692	6828	3521	0.6633	0.5989	0.7432